Work in Progress Report: Nonvisual Visual Programming

advertisement
Work in Progress Report:
Nonvisual Visual Programming
Clayton Lewis
Department of Computer Science
University of Colorado, Boulder
Clayton.Lewis@colorado.edu
Keywords: POP-I.A. Learning to program; POP-I.B. Barriers to programming; POP-III.C. Visual languages;
POP-IV.B. User interfaces
Abstract
Visual programming systems are widely used to introduce children and other learners to
programming, but they cannot be used by blind people. Inspired by the ideas of blind computer
scientist T.V.Raman, the Noodle system provides nonvisual access to a dataflow programming
system, a popular model for visual programming systems. This paper describes the design and
implementation of Noodle and some of its capabilities. The paper suggests that the same approach
used to develop Noodle could be applied to increase the accessibility of other visual programming
systems for learners with disabilities.
1. Visual programming systems are popular ways of introducing programming to
children and other learners.
Despite repeated arguments for their usefulness (e.g. Glinert, 1990; Victor, 2013) visual programming
systems have yet to find widespread adoption in a world still dominated by textual languages. But in
one important application such systems are now dominant: the introduction of computing to children.
Scratch, a visual adaptation of the textual Logo language, has millions of users (Resnick et al., 2009,
http://scratch.mit.edu/). Agentsheets, a rule-based language presented in a visual environment rather
than textually, is also widely used (Repenning and Ioannidou, 2004, http://www.agentsheets.com/).
LabView, a visual programming system originally developed for the control of electronic instruments,
is provided as the programming medium for the Lego Mindstorms robotics kits, including WeDO
(http://www.ni.com/academic/wedo/), aimed explicitly at younger children (ages 7-11).
These visual programming systems aim to exploit the capabilities of vision to support understanding
of program structure and function. Graphical representations and layouts are used to help users
recognize the parts of programs, and how they are connected. How well visual systems deliver on
these intentions can be questioned; Green, Petre, Blackwell, and others have analysed visual
programming systems, with mixed findings (see e.g. Green and Petre, 1996; Blackwell, et al. 2001).
Likely the popularity of these languages for children and other learners owes a good deal to the
relative simplicity and small scale of the programs typical created by learners, for which the strain on
representational systems is reduced.
2. Visual programming systems cannot be used by learners who cannot see.
As noted as early as 1990 (Glinert, 1990, p. 4) visual programming systems present a serious
challenge for users who cannot see, for obvious reasons. (In fact, visual programming systems also
present barriers to sighted users with motor limitations, because of their typical reliance on mouse
interaction. The work to be discussed in this paper addresses these barriers as well, but more modest
adaptations of existing systems could likely remove or at least greatly reduce these problems, by
permitting navigation of their user interfaces via keyboard commands.)
The limitations of these systems mean that blind children are excluded from learning activities that
use them. This is deplorable, since programming is an activity that is in other ways quite accessible to
PPIG, University of Sussex, 2014
www.ppig.org
2
blind people, because of the flexible access provided by machine-readable representations of textual
information, including programs. But, one might feel, if visual programming tools make programming
easier to learn for other children, surely it must be right to use them. Equally, one might feel, there
just isn't any way to deliver the benefits of "visual" representations to learners who cannot see. But
there is a way to attack this painful dilemma.
3. The Raman Principle suggests that no activities are intrinsically visual.
Blind computer scientist T V Raman has suggested (personal communication, 2009) that a way to
think about the visual system is as a way to answer queries against a spatial database. If you have an
alternate way to ask the queries and get the answers, you don't need the visual system. While Raman
has not expressed this principle, in this explicit form, in his writings, related ideas can be found in
Raman (1996) and Raman and Gries (1997). See also Lewis (2013).
A clear illustration of the Raman Principle is an adaptation of the Jawbreaker computer game, created
by Raman and Chen (n.d.). As discussed in Lewis (2013), playing Jawbreaker involves clicking on
coloured balls, making clumps of balls of the same colour go away, and earning points in a way that
depends on the size of the clumps. Anyone seeing the game would imagine that playing it is an
intrinsically visual activity, and that no one who cannot see the balls and their arrangement could
possibly play the game. But Raman, who is totally blind, can play the adapted version of the game.
To create that version, Raman and Chen analysed the questions that sighted players use their eyes to
answer, and added keyboard commands, with speech output, that allow players to ask these questions
and receive the answers. For example, the scoring system of the game rewards a player who
determines which colour of ball is the commonest when the game begins, and avoids eliminating any
balls of that colour until the end of the game. That strategy yields the largest possible clump for
removal with one click, an effect that dominates the scoring, since slightly larger clumps give much
larger scores. A sighted player would judge which colour is commonest by "eyeballing", if playing
casually, or by counting, if playing more carefully. A blind player in the adapted game does it by
typing "n", and listening to a spoken report, that there are 13 red balls, 15 blue ones, and so on. Other
commands and responses allow the blind player to learn what they have to know about the
arrangement of the balls in the game.
The Raman Principle does not assert that the mental processes associated with an activity (playing a
game, in the example) will be identical for different presentations. Nevertheless it suggests a way to
make an activity supported by visual representations possible for non-sighted users.
3.1 The Raman Principle suggests that any visual programming system could have a
nonvisual counterpart that is operable without vision.
The logic of the Jawbreaker example can be extended to visual programming systems. For any such
system, one can ask, "What information is a sighted user obtaining from the visual display?" One can
then provide a means for a non-sighted user to pose a corresponding question, and to receive an
answer, using nonvisual media.
4. Noodle is a nonvisual dataflow programming system.
In dataflow systems, one popular form of visual programming system, programs consist of functional
units, whose values (outputs) may be connected to the inputs of other functional units, and whose
inputs may be connected to the outputs of other functional units. Connections are shown visually as
lines or curves, representing paths along which data flow from one functional unit to another.
Programs are built by selecting available functional units from a palette, placing them in a
construction area, and connecting their inputs and outputs.
Noodle provides a system of keyboard commands with speech output that supports creation,
execution, and modification of dataflow programs, with visual presentation being optional. That is,
Noodle provides a simple diagrammatic representation of programs, mainly for expository purposes,
PPIG, University of Sussex, 2014
www.ppig.org
3
but it is intended to be used without any use of these diagrams. Noodle is implemented in JavaScript
as a Web application.
Noodle's design is based on a task analysis of dataflow programming, influenced by the Raman and
Chen work. Their Jawbreaker user interface supports three kinds of user actions: navigation actions,
that do not modify the state of the game, but move a point of focus from one location on the game
screen to another; queries, that also do not modify the state of the game, but provided particular
information to the user; and operations, actions that do modify the state of the game. Adapting this
framework to dataflow programming, the design for Noodle provides navigation actions that move a
focus between different parts of the programming environment, for example, between the palette and
the construction area (called a workspace in Noodle), or between functional units (called blocks in
Noodle) already placed in the workspace. Noodle also provides operations, including commands that
cause blocks to be placed in the workspace, that cause data paths (called edges in Noodle) to be
placed between blocks, that cause blocks or edges to be deleted, and so forth. The current design
includes no queries. Instead, all navigation actions produce a description of the current entity of focus
(as discussed later, it may be appropriate to add some queries to the design.) All of these descriptions,
and in fact all feedback provided to the user, are presented in audio form, using text-to-speech
conversion (using the speak.js package, https://github.com/kripken/speak.js).
Table 1 shows a trace of the creation and execution of a trivial Noodle program.
_________________________________________________________________________________
user types "n" [“new” command]
system says:type name of new function
user types "foo"
system says:creating new workspace foo
user types 
system says:palette noteC
user types / [select]
system says:adding block noteC after
externalIn
user types 
system says:palette noteC [see note in
caption]
user types 
system says:heading sounds
user types 
system says:heading sound operations
user types 
system says:palette amplify
user types 
system says:palette attenuate
user types 
system says:palette mixSounds
user types 
system says:palette concatSounds
user types 
system says:palette playSound
user types / [select]
system says:adding block playSound after
noteC
user types 
system says:palette playSound
user types 
system says:heading sound operations
user types 
system says:workspace foo playSound
user types 
system says:input connector buffer on
block playSound with no edge.
user types / [select]
system says:connector selected
user types 
system says:input connector buffer on
block playSound with no edge.
user types 
system says:workspace foo playSound
user types 
system says:workspace foo noteC
user types 
system says:output connector buffer on
block noteC with no edge.
user types / [select]
system says:adding edge from connector
buffer on block noteC to connector buffer
on block playSound
user types "r" [“run” command]
system plays tone for middle C
________________________________________________________________________________
Table 1. Transcript of interaction during creation and execution of a program that plays a tone. Note:
When an arrow key is pressed after a command other than a navigation command, the current
location is described but not changed, as an aid to orientation.
PPIG, University of Sussex, 2014
www.ppig.org
4
5. Design and Implementation of Noodle.
5.1 Noodle uses a model-view architecture, with a declarative representation of programs.
The foundation of Noodle is an interpreter for dataflow programs that accepts programs represented
as JSON (JavaScript Object Notation) strings. This declarative representation, or model, can be
viewed graphically, and in fact could be created and manipulated via mouse interaction, in early
versions of Noodle. This same representation can be inspected and manipulated via a completely
different user interface, the completely nonvisual interface just described, that uses only keyboard
interaction and audio output. Thus a single model can be the subject of multiple views.
This architecture permits multiple user interfaces to be explored easily. Work on Noodle has included
four different interfaces: a conventional visual interface, now available for output only, for exposition;
the completely nonvisual interface that is the main subject of this paper; an earlier nonvisual interface
that used a larger, more complex set of keyboard commands (described in Lewis, 2013); and a buttonoriented interface intended for use on phones (currently incomplete.)
5.2 Noodle uses a pseudospatialized navigation scheme to reduce the number of keyboard
commands.
The first version of Noodle used a large number of keyboard commands for navigation. For example,
separate commands were used to move up and down in the palette, to move up and down in the
program construction area, to move among the connectors on blocks, and to follow edges. These
operations required a total of eight commands.
To reduce this load (both from learning the commands, and from locating the commands on the
keyboard, during use) four generic spatial movement commands (using arrow keys) are defined in the
current Noodle user interface. Roughly, the horizontal movement keys move the focus between virtual
columns (no spatial layout is actually shown), and the vertical keys move the focus up and down the
columns.
The arrangement is "pseudospatial" rather than fully "spatial", not only because no spatial layout is
shown, but also because some uses of the keys do not correspond to spatial moves in any simple way.
For example, one could think of the palette and workspace as "columns", one to the left of the other,
so that one moves “right” or “left” between them, and “up” or “down” within them. But moving to the
"right" from a block in the workspace leads to the connectors available on that block for the
attachment of edges. Moving further "right" from a connector, if there is a edge attached, leads onto
that edge, and moving "right" again leads to the connector on the far end of the edge, which is on
some other block in the program construction area. This is plainly not actually "spatial", since moving
to the "right" can eventually bring one to a point "above" or "below" the starting point.
5.3 Following an edge shows a contrast between nonvisual and visual dataflow
programming.
A key operation in understanding a dataflow program is tracing the data paths that connect the blocks
in a program. In visual representations of dataflow programs this is done by tracing lines or curves
from one unit to another. While conceptually straightforward, this tracing is often quite difficult to do
in practice. Paths are generally not straight, and, worse, frequently cross one another. In complex
programs several paths may run in parallel, close together, making it easy to slip over from one path
to a different one, while tracing.
Noodle's navigation scheme eliminates this problem. Tracing an edge from one unit to another can be
done directly, by moving onto the relevant connector, thence onto the edge, and on to the connector
on the other end of the edge. The situation of other edges cannot interfere at all with this process.
PPIG, University of Sussex, 2014
www.ppig.org
5
6. Capabilities of Noodle
6.1 Noodle supports simple sound processing.
To learn to program one needs to know what programs are doing. Guzdial's (2003) work on media
computation suggests that good learning environments allow learners to create media content, such as
graphics and sound. For blind learners sound is an attractive choice, and for that reason Noodle
includes primitive support for sound processing. Available blocks include ones that produce common
musical notes, that mix and concatenate sounds, and that allow the frequency and length of sounds to
be under program control.
6.2 Noodle offers the potential of programming on small screen devices.
Current programming systems for phones require large screens. While at least one effort offers
application development in Android devices
(https://play.google.com/store/apps/details?id=com.aide.ui) the aim seems to be to support tablets
with at least modest screens, rather than phones with small screens. The Noodle user interface could
be supported with no screen at all, and work has been done on a version intended to be run as a
mobile Web app for phones.
7. Nonvisual versions of other visual programming systems should be developed.
Task analysis, as suggested by the Raman Principle, could be applied to the design of alternative user
interfaces for visual programming systems that use paradigms other than dataflow. Here are a few
illustrative suggestions.
Scratch uses visual cues to distinguish semantic categories and simplify syntactic constraints. These
visual cues are the shapes, like the shapes of jigsaw puzzle pieces, that allow a visual judgement to be
made that a given piece of program will or will not fit into a program under construction at a given
place. If the user tries to place a piece of program in a wrong place, it will not fit. A major reason for
the popularity of Scratch, as compared to the older Logo language, is that users do not have to learn
complex textual syntax rules, and are not vulnerable to typing errors that produce syntactically invalid
code. How could these benefits be secured for users who cannot see?
In a nonvisual version of Scratch, nonvisual navigation commands could traverse gaps in a program,
that is, places in a program under construction where new material could be added. From such a gap,
other nonvisual navigation commands could directly access palettes of compatible program elements.
That is, instead of the user identifying the shape of a gap, and then visually matching the shape of the
gap against the shapes of available units in part of the palette, the user could move directly to the
relevant section of the palette. Navigation in this nonvisual scheme could be easier than in the current
visual scheme, for all users. If this proved to be the case, it could be added to the current Scratch user
interface, rather than creating a different, alternative interface.
Another aspect of Scratch that is not workable for blind users is program assembly by visually guided
dragging. For example, given a sequence of commands, the user can drag a "repeat" unit up to them,
so that it deforms to fit around them, forming a loop with the sequence of commands as the body.
Nonvisual placement actions could be provided to support this action. A user would select the
beginning and end of the loop body, and then select the repeat unit, and the loop structure could be
completed automatically.
Most uses of Scratch today produce graphical, animated output, which would be problematic today
for blind learners. In time, support could be contrived for interpreting graphical output nonvisually, as
was suggested above for Noodle.
Like Scratch, Agentsheets uses separate palettes for program pieces that play different roles, for
example separating conditions and actions for rules (Agentsheets programs are collections of if-then
rules.) As for Scratch, nonvisual navigation commands could support this access.
PPIG, University of Sussex, 2014
www.ppig.org
6
As for Noodle and Scratch, supporting understanding of dynamic graphic behaviour, as well as static
arrangements of elements, is an important challenge. Agentsheets has "conversational programming"
features that provide (visual) explanations of some aspects of dynamic program behaviour, such as
which conditions of rules are currently satisfied. These features might be adapted to provide spoken
commentary on program execution, and this enhancement could be of value to sighted users, too.
More generally, other design features of nonvisual presentations, such as easier tracing of connections
between program elements, or easier access to relevant palette items from a program under
construction, could be added to existing visual languages.
8. Conclusion
Nonvisual visual programming is possible, and offers potential benefits to learners who cannot see. Its
development may also offer benefit to other users, as is common for work on inclusive design.
9. Acknowledgements
Thanks to Antranig Basman, Colin Clark, Greg Elin, Jamal Mazrui, T V Raman, and Gregg
Vanderheiden for useful discussions and encouragement, and to anonymous reviewers for helpful
suggestions.
10. References
Blackwell, A.F., Whitley, K.N., Good, J. and Petre, M. (2001). Cognitive factors in programming
with diagrams. Artificial Intelligence Review 15(1), 95-113.
Glinert, E. (ed.) (1990) Visual programming environments: Paradigms and systems. Los Alamitos,
CA: IEEE Computer Society Press.
Green, T. R. G., & Petre, M. (1996). Usability analysis of visual programming environments: a
‘cognitive dimensions’ framework. Journal of Visual Languages & Computing, 7(2), 131-174.
Guzdial, Mark (2003.)A media computation course for non-majors. SIGCSE Bull. 35, 3 (June 2003),
104-108.
Lewis, C. (2013) Pushing the Raman principle. In Proceedings of the 10th International CrossDisciplinary Conference on Web Accessibility (W4A '13). ACM, New York, NY, USA, Article
18, 4 pages.
Raman, T.V. (1996) Emacspeak-- A speech interface. In Proceedings of the SIGCHI Conference on
Human Factors in Computing Systems (CHI '96), Michael J. Tauber (Ed.). ACM, New York, NY,
USA, 66-71.
Raman, T.V. and Gries, D. (1997.) Documents mean more than just paper! Mathematical and
Computer Modelling, Volume 26, Issue 1, July, 45-53.
Raman, T.V. and Chen, C.L. (n.d. ) AxsJAX-Enhanced Jawbreaker User Guide. Retrieved from
http://google-axsjax.googlecode.com/svn-history/r540/trunk/docs/jawbreaker_userguide.
Repenning, A., & Ioannidou, A. (2004). Agent-based end-user development. Communications of the
ACM, 47(9), 43-46.
Repenning, A. (2013) Conversational programming: exploring interactive program analysis. In
Proceedings of the 2013 ACM international symposium on New ideas, new paradigms, and
reflections on programming & software (Onward! '13). ACM, New York, NY, USA, 63-74.
Resnick, M., Maloney, J., Monroy-Hernández, A., Rusk, N., Eastmond, E., Brennan, K., Millner, A.,
Rosenbaum, E., Silver, J., Silverman, B., and Kafai, Y. (2009) Scratch: programming for all.
Communications of the ACM 52, 11 (November 2009), 60-67.
Victor, B. (2013) The future of programming. Retrieved from
http://worrydream.com/#!/TheFutureOfProgramming
PPIG, University of Sussex, 2014
www.ppig.org
Download